In brief: attackers now hand most of the work to software and move faster than a person reading logs. A blue-team agent that lives next to your servers changes the odds for anyone without a security team. This post covers what we borrowed from immune cells, what we borrowed from bacteria (the better model, according to Lennart Lopin’s book The Oldest War), and what our own agent got wrong on its first day.
On its first day watching seven servers, our immune-system agent sent a dozen high-severity emails. Every one of them was about the same intruder: its own human partner, logging in as root.
Immunologists have a word for a defense that attacks the body it protects: autoimmunity. We’ll come back to how we cured it, because the cure turned out to be the most useful thing we learned. First, why anyone running a few servers needs an immune system at all.
The other side already automated
In November 2025 Anthropic reported that it had caught a Chinese state-sponsored group, which it calls GTG-1002, using Claude Code to run an espionage campaign against about thirty organizations. The AI did 80 to 90 percent of the tactical work on its own: reconnaissance, finding weaknesses, writing exploits, moving through networks. At its peak it made thousands of requests, often several per second, at rates Anthropic called “physically impossible” for people. The same report notes that the AI often overstated what it found and sometimes made things up. Machine speed isn’t machine accuracy, but it is speed.
The rest of the field shows the same curve. CrowdStrike measures breakout time, how long a criminal intruder needs to get from the first machine it compromised to the next one. It was 62 minutes on average in 2023, 48 in 2024 and 29 in 2025, and the fastest breakout CrowdStrike saw in 2025 took 27 seconds. Mandiant’s M-Trends 2026 puts the mean time to exploit a vulnerability at an estimated minus seven days: attacks routinely start before a patch exists. When one group breaks in and hands the access to another, that handoff took more than eight hours in 2022. In 2025 it took 22 seconds.
Red teams have the same tools. In June 2025 XBOW, an autonomous penetration tester, became the first to reach the top of HackerOne’s US bug bounty leaderboard. And a new server gets no grace period: Orca Security’s cloud honeypots saw their SSH port accessed within four minutes of going online.
Now picture the defender in most of these stories. It isn’t a security operations center with analysts on shifts. It’s one person with a handful of servers, a day job, and a habit of reading the logs when something breaks. That person is defending at human speed against tools that never sleep and never skip the boring checks.
You can’t fix that by asking the person to look more often. You fix it the way the body did: you put something on patrol.
What an immune cell does that a firewall doesn’t
A firewall is a wall. It checks what crosses a border and drops what matches a rule. That’s useful, and it’s also the oldest trick in biology. Bacteria evolved it billions of years ago (it’s called restriction-modification), and their viruses learned to slip past it about as fast.
An immune system works differently. It patrols. Cells move through tissue all day, sampling what they find and comparing it with what belongs there. A few things about how that works carried straight over into our agent.
It patrols on a schedule. Every hour, our immune-system template takes the vital signs of each server it watches: load, memory, disks, failed and accepted SSH logins, listening ports, pending security updates and TLS certificates. The deeper look comes from its blue-team skill: new users and keys, cron entries, processes running from /tmp, web shells, miners, reverse shells. Sentinel, the agent we run on our own servers, does that deeper check every morning and a hardening review every Monday.
It knows what belongs. The first time the agent visits a machine, it takes a baseline: what runs there, which ports are open, which users have shells, what the cron tables say. Every sweep after that is about drift: what’s new, what’s gone, what got worse. This is the immune system’s oldest question, self or not self, asked of a Linux box.
It carries evidence, and someone else decides. The closest immune cell to our agent is the dendritic cell. It samples tissue, picks up a fragment of whatever looks foreign, and carries it to a lymph node, where T cells decide whether to attack. Our agent does the same. It quotes the command and the lines of output that show a problem, says how sure it is, and gives the exact commands it would run. Then it stops. It logs in as a user that can read logs and nothing else: no sudo, no Docker group, a locked password. It proposes; its human decides.
That last choice is deliberate, and it comes from computing’s best-known autoimmune event. In July 2024 a faulty update to CrowdStrike’s Falcon sensor, which runs with full kernel privilege, crashed about 8.5 million Windows machines. The defense took down the host. A read-only agent can be wrong, but it can’t do that.
The first day: autoimmunity
Back to the dozen emails. Sentinel’s first day produced about fifty reports and about $3.35 of model tokens, and the “high” alerts were all its own human’s logins. The agent was doing its job and still getting it wrong. It saw root logins from addresses it didn’t know, on machines it had known for a few hours, and it had no way to tell its partner from an attacker.
The body solves this in the thymus. Young T cells that react to the body’s own tissue are killed there before they ever reach the bloodstream, so what survives can only react to what isn’t you. Sentinel got a thymus of its own, a short file listing the SSH keys that belong: whose they are and which machines they open. Logins are now judged by the key that was used, never by the address they came from. An unknown key is news. A known key from a new café is not.
The second fix was regulation. Your immune system has cells whose whole job is to calm the others down once a threat is handled, because inflammation that never stops does damage of its own. We gave Sentinel a plain rule, “quiet is the default”, and reworded its duties so it answers with silence unless something is new or worse. Findings live in one ledger, numbered, so the same open port isn’t reported as news every hour.
Neither fix was a smarter model. Both were about teaching the agent what self looks like, and when to stop talking.
Why the book says: think bacteria
Lennart Lopin, who builds bootkey, wrote a book about exactly this problem with Tana von Kepler: The Oldest War: Biology, Bacteria, and the Future of Cyber Defense. It opens with a warning about the comparison this post has been making.
In 1994 the computer scientist Stephanie Forrest proposed that a computer could tell self from non-self the way your immune system does: record what a healthy machine does, generate random detectors, delete the ones that match normal behavior, and keep the rest as sentinels. It launched a field, artificial immune systems, with conferences, theses and a startup. By the mid-2010s no production security tool of any weight used it. The detectors you need grow exponentially with what you watch, and in 2005 Thomas Stibor and colleagues showed that plain statistical anomaly detectors did the job better.
The book’s diagnosis is that Forrest copied the wrong organism. Your immune system learns by brute force: it makes something like a hundred million new lymphocytes a day, educates them in an organ built for the purpose, and lets them compete. Software has no equivalent. Bacteria, which have been fighting viruses for about four billion years, do something else. They carry defense islands: compact modules, each with a sensor that spots the invader and an effector that responds. And they pass these modules between cells. A defense that arrives in a new cell works there immediately, against attackers that cell has never met. Biologists knew of about ten such systems in 2017 and have cataloged more than 150 families since. In the book’s words, the defense of the AI era will look like “an ecosystem of modular, composable, horizontally distributable defense parts rather than a smarter antivirus.”
That describes an agent template better than we could have. Our immune-system template is one module: a sensor (the hourly sweep script), a library of known-bad patterns (the blue-team skill), a schedule (the duties) and a report format. It installs onto a fresh agent in about two minutes and starts working on its first visit, from a baseline instead of weeks of learning. Even Sentinel’s thymus is closer to the bacterial way than it looks. Nobody trained it; it’s a list of keys someone wrote down and handed over. And when one agent learns something worth keeping, like Sentinel’s quiet rules, that lesson can go back into the module, and every agent that boots with it afterwards starts there. That’s horizontal transfer, in plain files.
The book insists on saying where each analogy breaks, so here is where ours does. Bacterial effectors act: some kill their own cell so a virus can’t spread to its neighbors. Ours do nothing. A bacterium can lose cells by the billion; you can’t lose your one database server, and we’d rather an agent be slow to respond than able to break its host. The book’s picture of the security team of 2036 has the same split: “Humans set policy and agents execute.” For now our agents watch and recommend, and the human is the one who executes.
Machine speed, honestly
An hourly sweep is not machine speed, and we won’t pretend it is. If an intruder can move to the next machine in 27 seconds, an agent that looks once an hour will not catch the move.
Here is what it does catch. People who break into a server usually want to come back, so they leave something behind: a new key in authorized_keys, a cron entry that pipes curl into a shell, a second account with UID 0, a miner, a process running from /tmp, a web shell in an upload folder. That’s what the blue-team skill looks for on every machine, and it’s exactly what a busy person never checks.
Mandiant’s global median dwell time, from compromise to discovery, was 14 days in 2025. That figure comes from Mandiant’s own investigations, at organizations with someone paid to look. Nobody measures it for one person with five servers and no one looking. The agent doesn’t outrun the attacker. It closes the stretch of time when nobody is looking at all, from “whenever you next read the logs” down to the next sweep.
What it costs, and where it stops
The agent’s machine is $9 a month, billed by the hour, and it thinks on your own model key, so the tokens are yours at your provider’s rates. Sentinel’s noisy first day cost about $3.35 in tokens. A quiet sweep is mostly the model reading the script’s output and answering in one line.
Know its limits before you rely on it. It looks every hour, not every second, so it adds to a firewall, fail2ban and automatic updates and doesn’t replace any of them. It never changes a server: during an active compromise it leads with the evidence and gives you containment steps (cut the network, snapshot the disk before any reboot, rotate credentials), and you take them.
It also reads text written by strangers. Attackers pick their own usernames and file names, and all of that ends up in front of a model. That’s the main reason its logins can read and nothing more: if someone talks it into something, there’s little it can do. And it’s only as good as its model. A free model route may log your prompts and a weak model misses things, so pick one that scores well on security work and doesn’t log.
Try it
The immune-system template is the first choice in bootkey’s boot wizard. You pick a city and a model, add the servers it should watch, and give each one a read-only login with the agent’s public key. The guide walks you through it, or your own AI can do the setup with a one-hour key, the way Claude did for Sentinel. Boot an agent.
Sources
- Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign,” November 2025, and the announcement of November 13, 2025. GTG-1002, the 80 to 90 percent of tactical work, the request rates, about thirty targets, and the overstated findings.
- CrowdStrike, “2026 Global Threat Report,” February 24, 2026. Average eCrime breakout time of 29 minutes in 2025, fastest 27 seconds. Earlier years: the 2025 report (48 minutes) and the 2024 report (62 minutes).
- Mandiant, “M-Trends 2026,” Google Cloud, March 23, 2026. Mean time to exploit of an estimated minus seven days, the handoff window of 22 seconds, and the global median dwell time of 14 days.
- XBOW, “The road to Top 1: how XBOW did it,” June 24, 2025. The first autonomous penetration tester at the top of HackerOne’s US leaderboard.
- Orca Security, “2023 Honeypotting in the Cloud Report,” June 2023. SSH honeypots accessed within four minutes.
- Microsoft, “Helping our customers through the CrowdStrike outage,” July 20, 2024. About 8.5 million Windows devices affected.
- Lennart Lopin and Tana von Kepler, The Oldest War: Biology, Bacteria, and the Future of Cyber Defense, 2026, ISBN 979-8-197-63595-2. The history of artificial immune systems, defense islands and horizontal transfer, the catalog’s growth from about ten systems to more than 150 families, and the two quotations.
- S. Forrest, A. S. Perelson, L. Allen and R. Cherukuri, “Self-Nonself Discrimination in a Computer,” Proceedings of the 1994 IEEE Symposium on Security and Privacy, pp. 202-212. T. Stibor, J. Timmis and C. Eckert, “A Comparative Study of Real-Valued Negative Selection to Statistical Anomaly Detection Techniques,” ICARIS 2005.
- Sentinel’s first day: our own agent’s reports and token bill, November 13 and 14, 2026. The setup itself is described in Claude’s guest post.
Cite this page
bootkey. “Attackers run at machine speed now. Your servers need an immune system.” Updated November 15, 2026. https://bootkey.ai/research/immune-system